Adaptive multi-scale threshold with time-gated evaluation for real-time Fault detection in non-stationary and high-dynamic industrial cyber-physical systems

Abstract

Fault detection and diagnosis approaches play a vital role in maintaining reliability and resilience of industrial cyber-physical systems, especially in smart manufacturing settings characterized by complex interdependencies and dynamic operating conditions. Conventional threshold-based fault detection and diagnosis methods, including global or local, often struggle to detect faults reliably under non-stationary conditions, where current amplitudes and cycle durations are varying due to changes in load or machine behavior. To address these limitations, this article proposes a novel multi-scale fault detection approach that integrates global and local detection methods via an adaptive dual-threshold strategy. The proposed approach integrates exponentially weighted moving average for capturing global signal drifts with a peak-to-peak envelope for detecting localized deviations. A key innovation lies in the use of adaptive temporal scaling, where the window size used in local thresholding adjusts dynamically in real time based on signal variance. Allowing robust detection across both abrupt and incipient faults with high temporal accuracy. Additionally, the approach also employs station-specific thresholds and a time-gated evaluation mechanism to suppress false positives and ensure operational relevance and improve interpretability. The proposed approach is validated on a real-world conveyor-based sorting system, demonstrating superior performance across several fault types across multiple stations. Experimental comparative results demonstrate its effectiveness in multi-scale responsiveness while reducing false alarm rates compared with global and local methods. Detection rates exceeding 98%, zero false alarms, and robust accuracy under varying operational conditions. This multi-scale approach offers a scalable and deployable solution for real-time fault detection in next-generation industrial cyber-physical systems.